Reproductions and technical breakdowns of the latest papers in magnetic indoor positioning — each issue takes a close read of one representative work, analyzing its core idea, system architecture and engineering path to deployment, continuously fused and iterated with Gear-Lab's magnetic-inertial PDR solution.
Magnetic fingerprint positioning exploits the inherent distortion of the Earth's magnetic field by a building's steel structure to enable infrastructure-free localization. But raw three-axis magnetometer readings (Mx, My, Mz) are vectors in the device coordinate frame — once the device pose changes (portrait vs. landscape, pocket orientation, robot mounting angle), the fingerprint at the same location no longer matches at all. The classic workaround demands "mapping and positioning with consistent pose", which is nearly impossible to satisfy in real use. The MagNet study gives a quantitative conclusion: with 3D raw input under a fixed 90° rotation, the mean absolute error degrades from about 0.3 m to 13–14 m (roughly 40×), and this holds for every rotation axis.
Starting from the symmetry of the rotation group, MagNet compresses the 3D magnetic field vector into two rotation-invariant scalar features — discarding one dimension of information in exchange for complete immunity to device rotation.
Features are rotation-normalized before entering the network — robustness is not left for the network to "learn". No matter how the device rotates, the input representation is identical and the error curve stays flat regardless of rotation angle
In large buildings (strong fingerprint uniqueness) 2D features win even at zero rotation (Loomis: 1.46 m vs 1.96 m); in small-to-medium buildings there is a 5–6° crossover threshold — real handheld use almost certainly exceeds it
No pose constraints or alignment calibration needed between mapping and positioning; under crowdsourced mapping (random rotations in both training and test data) 2D features remain stable, directly enabling low-cost large-scale collection
A 7-layer dilated 1D convolutional network takes a magnetic feature sequence of 200 samples (50 Hz × 4 s sliding window) and directly regresses (x, y) coordinates — no initial position estimate needed, natively solving the "lost robot" problem (unlike CNN+RNN hybrid schemes).
MagPie is a public magnetic positioning benchmark released by the University of Illinois Urbana-Champaign (UIUC): 3 buildings (large/medium/small), handheld and robot platforms, ~689 ㎡ coverage, 50 Hz magnetometer + 200 Hz IMU, with centimeter-level ground truth. MagNet compared against prior SOTA systems on this benchmark:
| Building (Size) | CNN+RNN (IPIN 2021) | Prior Work (ICCAI 2025) | MagNetS (360k) | MagNetXL (1M) |
|---|---|---|---|---|
| CSL (small) | 0.30 m | 0.22 m | 0.22 m | 0.21 m |
| Talbot (medium) | 1.06 m | 0.82 m | 0.66 m | 0.64 m |
| Loomis (large) | 1.07 m | 1.98 m | 2.13 m | 1.96 m |
Mean absolute error MAE (meters), original MagPie no-rotation test set · green = best in column
Under a fixed 90° rotation (simulating a pocket-carry pose change) MAE rises from ~0.3 m to 13–14 m, with no difference across the X/Y/Z axes; under random rotation (simulating natural hand sway) the error grows monotonically with the standard deviation, up to ~11 m
With (Mn, Mg) input, MAE stays unchanged at any rotation magnitude; past the crossover threshold 2D input beats 3D across the board — large buildings 0°, medium 5°, small 6°, while real handheld use involves far more rotation
MagNetS achieves nearly identical accuracy with ~360k parameters (about 1/3 of MagNetXL) — the key version for mobile:
After INT8 quantization the model is ~0.4 MB in size; single-window inference is millisecond-scale 1D convolution, running in real time on phone NPU/DSP or wearable MCUs, fused online with PDR dead reckoning
Requiring no initial position estimate directly solves the "lost robot" relocalization problem; complementary to wheel odometry, and magnetic fingerprints are especially distinctive in warehouses packed with metal shelving
Positioning needs only one magnetometer + accelerometer — natively compatible with Gear-Lab's existing passive positioning terminals (BLE/LoRa/4G backhaul); magnetic maps can be updated at low cost via crowdsourced collection
Core literature and datasets cited in this topic. Note: as of 2026-09 the MagNet paper has no official code release; MagPie provides an official dataset and a community forum repository, so the architecture above can be reproduced directly with PyTorch/TensorFlow.
Frontier feature engineering and lightweight CNNs fuse seamlessly with Gear-Lab's magnetic-inertial PDR solution — we welcome technical evaluations and joint validation
Explore the PDR Positioning Solution → Contact UsLive subscription to the MQTT data streams of the proof-of-concept unit: step (BHI260AP PDR solved at 50 Hz, absolute coordinates reported on every step) and mag (QMC5883P sampled at 10 Hz / uploaded at 3 Hz, magnetic field in µT after ellipsoid calibration). Each step automatically pairs with the magnetic scalar |m| closest in time, mapped as a heatmap color onto the walking trajectory — magnetically anomalous zones (steel structures / equipment interference) show up directly on the track.